ja-style-check
Write or revise Japanese prose against a strict style standard.
Modes
report: return findings and suggestions, change no files.fix: edit files. Apply everything under Model-Applied Fixes in references/fix-policy.md even when it takes judgment, and report the rest.- No mode given: use
fix on an existing file. To write new Japanese prose, apply the rubric while drafting.
When the user wants new prose rather than a file edit, read references/rubric.md first, write directly in the requested format, and self-review against the rubric two or three times before answering.
Target files
Use the path the user gives. With no path, use the modified Markdown files. If neither identifies a target, ask for the file or the text.
Required reading
references/rubric.md before judging prose or writing Japanese.references/fix-policy.md before editing a file.references/output-schema.md before returning a report.
File workflow
- Settle the mode and the target files.
- Run the scanner. It reports candidates and never edits anything.
bash skills/ja-style-check/scripts/scan.sh [file...]
- Apply the deterministic fixes yourself with Edit, including the line joins.
mechanical-wrap comes marked auto_fixable; line-break does not, so read that pair before joining it. When you join two lines, add a space only if Latin script ends the first and begins the second; a Japanese join takes no separator.
- Judge the context-dependent findings using the scanner output and
references/rubric.md.
- Manually scan every heading, list item, blockquote, table row, and URL line for brackets (Check #2) and vague words (Check #4) — the scanner skips these lines entirely, so a short finding list is not evidence they are clean.
- In
fix mode, edit every Model-Applied Fix per references/fix-policy.md's rule for each: brackets, Latin script in Japanese sentences, meta prose, passive voice, jargon, the code-identifier rule (rewrite away an assignment-notation identifier, keep a traceability reference), paragraph and section structure, weak argument, reader load, voice, and notation. Report the Automatic Fixes Forbidden items — vague thresholds, absent first-use definitions, a hidden actor's real identity — as manual suggestions.
- Re-run the scanner after all edits above (steps 3, 5, and 6) to confirm none of them introduced a new candidate. A candidate you already judged and intentionally left as-is (such as a kept traceability identifier) reappearing is not "new" — stop once a run surfaces nothing beyond already-judged candidates, and never exceed three runs total.
- Return the result in the shape
references/output-schema.md defines.
Compression
When the user also asks to shorten the text, finish this skill first, then hand the result to extreme-compact. That order is deliberate: unfolding brackets and translating Latin words lengthens the text, so compressing first only re-introduces what this skill removes.
Non-negotiable
- Never guess a real-world fact the text doesn't state: a vague word's threshold, a term's definition, a hidden actor's identity. When a fix could change meaning, report it instead of editing (the Safety Rule in
references/fix-policy.md).
- 「確認して」 and 「チェックして」 alone are not requests for style fixes.
- Never rely on the scanner alone. It only finds candidates; the agent makes the style call.
- The scanner reports
brackets as high without knowing whether an exception applies. Decide per rubric Check #2 whether the bracket is a first-use definition, or Markdown syntax such as a link target, and clear the finding when it is.
- Latin script in Japanese prose has no scanner rule. Check #8 is yours to apply by reading. A deterministic version was attempted and withdrawn: telling a Japanese sentence containing English from an English sentence quoting Japanese kept producing missed violations, and this repo's own skill bodies are now English with Japanese examples, so the false positives landed on correct text. Read for it; do not expect a finding.
- The scanner sees paragraph text only. Headings, list items, blockquotes, table rows, and any line carrying a URL are skipped by every rule, so read those lines yourself rather than treating a short finding list as a clean bill of health.
- The scanner never writes. It has no
--fix; the flag is accepted and ignored. Every edit goes through Edit so you can see and review it.
1---2name: ja-style-check3description: Use this skill for Japanese writing and Japanese text revision unless the user only asks to inspect unrelated facts. Always trigger for requests to draft, rewrite, polish, summarize, or produce Japanese prose, including SKILL.md, PR descriptions, design docs, specs, README files, reviews, comments, and agent instructions. Holds Japanese output to near-zero parentheses and no bare Latin words inside a Japanese sentence. 「日本語で書いて」「文章を修正して」「自然な日本語にして」「要約して」「括弧を減らして」「読みやすくして」「文体を整えて」でも起動。 Do not trigger merely because the user says 「確認して」 or 「チェックして」 without asking for Japanese writing or style fixes.4---56# ja-style-check78Write or revise Japanese prose against a strict style standard.910## Modes1112- `report`: return findings and suggestions, change no files.13- `fix`: edit files. Apply everything under Model-Applied Fixes in `references/fix-policy.md` even when it takes judgment, and report the rest.14- No mode given: use `fix` on an existing file. To write new Japanese prose, apply the rubric while drafting.1516When the user wants new prose rather than a file edit, read `references/rubric.md` first, write directly in the requested format, and self-review against the rubric two or three times before answering.1718## Target files1920Use the path the user gives. With no path, use the modified Markdown files. If neither identifies a target, ask for the file or the text.2122## Required reading2324- `references/rubric.md` before judging prose or writing Japanese.25- `references/fix-policy.md` before editing a file.26- `references/output-schema.md` before returning a report.2728## File workflow29301. Settle the mode and the target files.312. Run the scanner. It reports candidates and never edits anything.3233```bash34bash skills/ja-style-check/scripts/scan.sh [file...]35```36373. Apply the deterministic fixes yourself with Edit, including the line joins. `mechanical-wrap` comes marked `auto_fixable`; `line-break` does not, so read that pair before joining it. When you join two lines, add a space only if Latin script ends the first and begins the second; a Japanese join takes no separator.384. Judge the context-dependent findings using the scanner output and `references/rubric.md`.395. Manually scan every heading, list item, blockquote, table row, and URL line for brackets (Check #2) and vague words (Check #4) — the scanner skips these lines entirely, so a short finding list is not evidence they are clean.406. In `fix` mode, edit every Model-Applied Fix per `references/fix-policy.md`'s rule for each: brackets, Latin script in Japanese sentences, meta prose, passive voice, jargon, the code-identifier rule (rewrite away an assignment-notation identifier, keep a traceability reference), paragraph and section structure, weak argument, reader load, voice, and notation. Report the Automatic Fixes Forbidden items — vague thresholds, absent first-use definitions, a hidden actor's real identity — as manual suggestions.417. Re-run the scanner after all edits above (steps 3, 5, and 6) to confirm none of them introduced a new candidate. A candidate you already judged and intentionally left as-is (such as a kept traceability identifier) reappearing is not "new" — stop once a run surfaces nothing beyond already-judged candidates, and never exceed three runs total.428. Return the result in the shape `references/output-schema.md` defines.4344## Compression4546When the user also asks to shorten the text, finish this skill first, then hand the result to extreme-compact. That order is deliberate: unfolding brackets and translating Latin words lengthens the text, so compressing first only re-introduces what this skill removes.4748## Non-negotiable4950- Never guess a real-world fact the text doesn't state: a vague word's threshold, a term's definition, a hidden actor's identity. When a fix could change meaning, report it instead of editing (the Safety Rule in `references/fix-policy.md`).51- 「確認して」 and 「チェックして」 alone are not requests for style fixes.52- Never rely on the scanner alone. It only finds candidates; the agent makes the style call.53- The scanner reports `brackets` as `high` without knowing whether an exception applies. Decide per rubric Check #2 whether the bracket is a first-use definition, or Markdown syntax such as a link target, and clear the finding when it is.54- **Latin script in Japanese prose has no scanner rule.** Check #8 is yours to apply by reading. A deterministic version was attempted and withdrawn: telling a Japanese sentence containing English from an English sentence quoting Japanese kept producing missed violations, and this repo's own skill bodies are now English with Japanese examples, so the false positives landed on correct text. Read for it; do not expect a finding.55- The scanner sees paragraph text only. Headings, list items, blockquotes, table rows, and any line carrying a URL are skipped by every rule, so read those lines yourself rather than treating a short finding list as a clean bill of health.56- The scanner never writes. It has no `--fix`; the flag is accepted and ignored. Every edit goes through Edit so you can see and review it.
maro114510